<p>I’ve written up how I build AI skills at incident.io, using our daily AI spend report skill as the worked example. Claude runs it every morning to analyse our inference costs and flag anything anomalous.</p>

<p>The post covers:</p>

<ul>
  <li>Solving the underlying problem with an agent first, before writing any skill (otherwise you encode whatever assumptions you were carrying)</li>
  <li>Splitting reference content into the codebase so the skill stays a thin runbook</li>
  <li>Using a fresh sub-agent to test the skill against an ambiguous/wrong/improveable feedback schema</li>
  <li>Handing the iteration loop to the agent itself once you have a clear success criterion</li>
</ul>

<p>The framing draws on Atul Gawande’s The Checklist Manifesto, which I think is the best treatment of writing this kind of executable process documentation.</p>

<p>Link in 👇</p>